add texture micro and macro tiling to radeon/r200 driver. This can improve performance up to 15% in texture-intensive applications. Convert the driver to use the correct blit format and blit width instead of fixed blit format and blit width when uploading textures to make it work.

1 files changed, 56 insertions, 14 deletions

diff --git a/src/mesa/drivers/dri/radeon/radeon_texstate.c b/src/mesa/drivers/dri/radeon/radeon_texstate.c
index 5e818da9fd1..b96ad740d15 100644
--- a/src/mesa/drivers/dri/radeon/radeon_texstate.c
+++ b/src/mesa/drivers/dri/radeon/radeon_texstate.c
@@ -127,8 +127,8 @@ static void radeonSetTexImages( radeonContextPtr rmesa,
 {
    radeonTexObjPtr t = (radeonTexObjPtr)tObj->DriverData;
    const struct gl_texture_image *baseImage = tObj->Image[0][tObj->BaseLevel];
-   GLint curOffset;
-   GLint i;
+   GLint curOffset, blitWidth;
+   GLint i, texelBytes;
    GLint numLevels;
    GLint log2Width, log2Height, log2Depth;
 
@@ -148,6 +148,7 @@ static void radeonSetTexImages( radeonContextPtr rmesa,
       return;
    }
 
+   texelBytes = baseImage->TexFormat->TexelBytes;
 
    /* Compute which mipmap levels we really want to send to the hardware.
     */
@@ -166,6 +167,34 @@ static void radeonSetTexImages( radeonContextPtr rmesa,
     * memory organized as a rectangle of width BLIT_WIDTH_BYTES.
     */
    curOffset = 0;
+   blitWidth = BLIT_WIDTH_BYTES;
+   t->tile_bits = 0;
+
+   /* figure out if this texture is suitable for tiling. */
+   if (texelBytes && (tObj->Target != GL_TEXTURE_RECTANGLE_NV)) {
+      if (rmesa->texmicrotile && (baseImage->Height > 1)) {
+	 /* allow 32 (bytes) x 1 mip (which will use two times the space
+	    the non-tiled version would use) max if base texture is large enough */
+	 if ((numLevels == 1) ||
+	   (((baseImage->Width * texelBytes / baseImage->Height) <= 32) &&
+	       (baseImage->Width * texelBytes > 64)) ||
+	    ((baseImage->Width * texelBytes / baseImage->Height) <= 16)) {
+	    /* R100 has two microtile bits (only the txoffset reg, not the blitter)
+	       weird: X2 + OPT: 32bit correct, 16bit completely hosed
+		      X2: 32bit correct, 16bit correct
+		      OPT: 32bit large mips correct, small mips hosed, 16bit completely hosed */
+	    t->tile_bits |= RADEON_TXO_MICRO_TILE_X2 /*| RADEON_TXO_MICRO_TILE_OPT*/;
+	 }
+      }
+      if ((baseImage->Width * texelBytes >= 256) && (baseImage->Height >= 16)) {
+	 /* R100 disables macro tiling only if mip width is smaller than 256 bytes, and not
+	    in the case if height is smaller than 16 (not 100% sure), as does the r200,
+	    so need to disable macro tiling in that case */
+	 if ((numLevels == 1) || ((baseImage->Width * texelBytes / baseImage->Height) <= 4)) {
+	    t->tile_bits |= RADEON_TXO_MACRO_TILE;
+	 }
+      }
+   }
 
    for (i = 0; i < numLevels; i++) {
       const struct gl_texture_image *texImage;
@@ -197,28 +226,41 @@ static void radeonSetTexImages( radeonContextPtr rmesa,
             else size = texImage->CompressedSize;
       }
       else if (tObj->Target == GL_TEXTURE_RECTANGLE_NV) {
-      	 size = ((texImage->Width * texImage->TexFormat->TexelBytes + 63)
-      	         & ~63) * texImage->Height;
+	 size = ((texImage->Width * texelBytes + 63) & ~63) * texImage->Height;
+      }
+      else if (t->tile_bits & RADEON_TXO_MICRO_TILE_X2) {
+	 /* tile pattern is 16 bytes x2. mipmaps stay 32 byte aligned,
+	    though the actual offset may be different (if texture is less than
+	    32 bytes width) to the untiled case */
+	 int w = (texImage->Width * texelBytes * 2 + 31) & ~31;
+	 size = (w * ((texImage->Height + 1) / 2)) * texImage->Depth;
+	 blitWidth = MAX2(texImage->Width, 64 / texelBytes);
       }
       else {
-         int w = texImage->Width * texImage->TexFormat->TexelBytes;
-         if (w < 32)
-            w = 32;
-         size = w * texImage->Height * texImage->Depth;
+	 int w = (texImage->Width * texelBytes + 31) & ~31;
+	 size = w * texImage->Height * texImage->Depth;
+	 blitWidth = MAX2(texImage->Width, 64 / texelBytes);
       }
       assert(size > 0);
 
-
       /* Align to 32-byte offset.  It is faster to do this unconditionally
        * (no branch penalty).
        */
 
       curOffset = (curOffset + 0x1f) & ~0x1f;
 
-      t->image[0][i].x = curOffset % BLIT_WIDTH_BYTES;
-      t->image[0][i].y = curOffset / BLIT_WIDTH_BYTES;
-      t->image[0][i].width  = MIN2(size, BLIT_WIDTH_BYTES);
-      t->image[0][i].height = size / t->image[0][i].width;
+      if (texelBytes) {
+	 t->image[0][i].x = curOffset; /* fix x and y coords up later together with offset */
+	 t->image[0][i].y = 0;
+	 t->image[0][i].width = MIN2(size / texelBytes, blitWidth);
+	 t->image[0][i].height = (size / texelBytes) / t->image[0][i].width;
+      }
+      else {
+         t->image[0][i].x = curOffset % BLIT_WIDTH_BYTES;
+         t->image[0][i].y = curOffset / BLIT_WIDTH_BYTES;
+         t->image[0][i].width  = MIN2(size, BLIT_WIDTH_BYTES);
+         t->image[0][i].height = size / t->image[0][i].width;     
+      }
 
 #if 0
       /* for debugging only and only  applicable to non-rectangle targets */
@@ -263,7 +305,7 @@ static void radeonSetTexImages( radeonContextPtr rmesa,
    if (baseImage->IsCompressed)
       t->pp_txpitch = (tObj->Image[0][t->base.firstLevel]->Width + 63) & ~(63);
    else
-      t->pp_txpitch = ((tObj->Image[0][t->base.firstLevel]->Width * baseImage->TexFormat->TexelBytes) + 63) & ~(63);
+      t->pp_txpitch = ((tObj->Image[0][t->base.firstLevel]->Width * texelBytes) + 63) & ~(63);
    t->pp_txpitch -= 32;
 
    t->dirty_state = TEX_ALL;